This is for your own bots/agents, or if you are a builder making them for your users.
Onboard first
A good bot/agent doesn’t ask you for something it can find. Before it starts talking, it needs to know relevant things about the user.
Start with what’s already public. Name, site, GitHub, the posts, the company page. None of that needs a permission prompt.
Then ask before you touch the private stuff. Email, the other accounts, sibling bots if they exist. Pull what you can once they say yes. Save it.
Then ask only the questions that are still missing. Ask user to voice-dictate it so you get a detailed download.
Now, set up plugins
Once that’s written down, propose and set up the plugins/connectors relevant for the agent’s tasks. Ideally any of these: Official plugin > CLI with a skill > API with skill > MCP > Browser
Loops
Then propose relevant routines/loops for the user.
- Time-based: every day, twice a week, whatever is best for.
- And/or a webhook if something should fire it - meeting booked, sentry alert, linear change, github trigger, etc.
This is a great start
But you still have other problems to solve: long term memory, compaction, model routing, caching and how you do evals.
I’ll write about these sometime too. Feel free to send any feedback about this article.